mm/zswap.c

   1 /*
   2  * zswap.c - zswap driver file
   3  *
   4  * zswap is a backend for frontswap that takes pages that are in the process
   5  * of being swapped out and attempts to compress and store them in a
   6  * RAM-based memory pool.  This can result in a significant I/O reduction on
   7  * the swap device and, in the case where decompressing from RAM is faster
   8  * than reading from the swap device, can also improve workload performance.
   9  *
  10  * Copyright (C) 2012  Seth Jennings <sjenning@linux.vnet.ibm.com>
  11  *
  12  * This program is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU General Public License
  14  * as published by the Free Software Foundation; either version 2
  15  * of the License, or (at your option) any later version.
  16  *
  17  * This program is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  * GNU General Public License for more details.
  21 */
  22
  23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  24
  25 #include <linux/module.h>
  26 #include <linux/cpu.h>
  27 #include <linux/highmem.h>
  28 #include <linux/slab.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/types.h>
  31 #include <linux/atomic.h>
  32 #include <linux/frontswap.h>
  33 #include <linux/rbtree.h>
  34 #include <linux/swap.h>
  35 #include <linux/crypto.h>
  36 #include <linux/mempool.h>
  37 #include <linux/zpool.h>
  38
  39 #include <linux/mm_types.h>
  40 #include <linux/page-flags.h>
  41 #include <linux/swapops.h>
  42 #include <linux/writeback.h>
  43 #include <linux/pagemap.h>
  44
  45 /*********************************
  46 * statistics
  47 **********************************/
  48 /* Total bytes used by the compressed storage */
  49 static u64 zswap_pool_total_size;
  50 /* The number of compressed pages currently stored in zswap */
  51 static atomic_t zswap_stored_pages = ATOMIC_INIT(0);
  52
  53 /*
  54  * The statistics below are not protected from concurrent access for
  55  * performance reasons so they may not be a 100% accurate.  However,
  56  * they do provide useful information on roughly how many times a
  57  * certain event is occurring.
  58 */
  59
  60 /* Pool limit was hit (see zswap_max_pool_percent) */
  61 static u64 zswap_pool_limit_hit;
  62 /* Pages written back when pool limit was reached */
  63 static u64 zswap_written_back_pages;
  64 /* Store failed due to a reclaim failure after pool limit was reached */
  65 static u64 zswap_reject_reclaim_fail;
  66 /* Compressed page was too big for the allocator to (optimally) store */
  67 static u64 zswap_reject_compress_poor;
  68 /* Store failed because underlying allocator could not get memory */
  69 static u64 zswap_reject_alloc_fail;
  70 /* Store failed because the entry metadata could not be allocated (rare) */
  71 static u64 zswap_reject_kmemcache_fail;
  72 /* Duplicate store was encountered (rare) */
  73 static u64 zswap_duplicate_entry;
  74
  75 /*********************************
  76 * tunables
  77 **********************************/
  78
  79 /* Enable/disable zswap (disabled by default) */
  80 static bool zswap_enabled;
  81 static int zswap_enabled_param_set(const char *,
  82                                    const struct kernel_param *);
  83 static struct kernel_param_ops zswap_enabled_param_ops = {
  84         .set =          zswap_enabled_param_set,
  85         .get =          param_get_bool,
  86 };
  87 module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
  88
  89 /* Crypto compressor to use */
  90 #define ZSWAP_COMPRESSOR_DEFAULT "lzo"
  91 static char *zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
  92 static int zswap_compressor_param_set(const char *,
  93                                       const struct kernel_param *);
  94 static struct kernel_param_ops zswap_compressor_param_ops = {
  95         .set =          zswap_compressor_param_set,
  96         .get =          param_get_charp,
  97         .free =         param_free_charp,
  98 };
  99 module_param_cb(compressor, &zswap_compressor_param_ops,
 100                 &zswap_compressor, 0644);
 101
 102 /* Compressed storage zpool to use */
 103 #define ZSWAP_ZPOOL_DEFAULT "zbud"
 104 static char *zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
 105 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
 106 static struct kernel_param_ops zswap_zpool_param_ops = {
 107         .set =          zswap_zpool_param_set,
 108         .get =          param_get_charp,
 109         .free =         param_free_charp,
 110 };
 111 module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
 112
 113 /* The maximum percentage of memory that the compressed pool can occupy */
 114 static unsigned int zswap_max_pool_percent = 20;
 115 module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
 116
 117 /*********************************
 118 * data structures
 119 **********************************/
 120
 121 struct zswap_pool {
 122         struct zpool *zpool;
 123         struct crypto_comp * __percpu *tfm;
 124         struct kref kref;
 125         struct list_head list;
 126         struct work_struct work;
 127         struct hlist_node node;
 128         char tfm_name[CRYPTO_MAX_ALG_NAME];
 129 };
 130
 131 /*
 132  * struct zswap_entry
 133  *
 134  * This structure contains the metadata for tracking a single compressed
 135  * page within zswap.
 136  *
 137  * rbnode - links the entry into red-black tree for the appropriate swap type
 138  * offset - the swap offset for the entry.  Index into the red-black tree.
 139  * refcount - the number of outstanding reference to the entry. This is needed
 140  *            to protect against premature freeing of the entry by code
 141  *            concurrent calls to load, invalidate, and writeback.  The lock
 142  *            for the zswap_tree structure that contains the entry must
 143  *            be held while changing the refcount.  Since the lock must
 144  *            be held, there is no reason to also make refcount atomic.
 145  * length - the length in bytes of the compressed page data.  Needed during
 146  *          decompression
 147  * pool - the zswap_pool the entry's data is in
 148  * handle - zpool allocation handle that stores the compressed page data
 149  */
 150 struct zswap_entry {
 151         struct rb_node rbnode;
 152         pgoff_t offset;
 153         int refcount;
 154         unsigned int length;
 155         struct zswap_pool *pool;
 156         unsigned long handle;
 157 };
 158
 159 struct zswap_header {
 160         swp_entry_t swpentry;
 161 };
 162
 163 /*
 164  * The tree lock in the zswap_tree struct protects a few things:
 165  * - the rbtree
 166  * - the refcount field of each entry in the tree
 167  */
 168 struct zswap_tree {
 169         struct rb_root rbroot;
 170         spinlock_t lock;
 171 };
 172
 173 static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
 174
 175 /* RCU-protected iteration */
 176 static LIST_HEAD(zswap_pools);
 177 /* protects zswap_pools list modification */
 178 static DEFINE_SPINLOCK(zswap_pools_lock);
 179 /* pool counter to provide unique names to zpool */
 180 static atomic_t zswap_pools_count = ATOMIC_INIT(0);
 181
 182 /* used by param callback function */
 183 static bool zswap_init_started;
 184
 185 /* fatal error during init */
 186 static bool zswap_init_failed;
 187
 188 /*********************************
 189 * helpers and fwd declarations
 190 **********************************/
 191
 192 #define zswap_pool_debug(msg, p)                                \
 193         pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,         \
 194                  zpool_get_type((p)->zpool))
 195
 196 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
 197 static int zswap_pool_get(struct zswap_pool *pool);
 198 static void zswap_pool_put(struct zswap_pool *pool);
 199
 200 static const struct zpool_ops zswap_zpool_ops = {
 201         .evict = zswap_writeback_entry
 202 };
 203
 204 static bool zswap_is_full(void)
 205 {
 206         return totalram_pages * zswap_max_pool_percent / 100 <
 207                 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 208 }
 209
 210 static void zswap_update_total_size(void)
 211 {
 212         struct zswap_pool *pool;
 213         u64 total = 0;
 214
 215         rcu_read_lock();
 216
 217         list_for_each_entry_rcu(pool, &zswap_pools, list)
 218                 total += zpool_get_total_size(pool->zpool);
 219
 220         rcu_read_unlock();
 221
 222         zswap_pool_total_size = total;
 223 }
 224
 225 /*********************************
 226 * zswap entry functions
 227 **********************************/
 228 static struct kmem_cache *zswap_entry_cache;
 229
 230 static int __init zswap_entry_cache_create(void)
 231 {
 232         zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
 233         return zswap_entry_cache == NULL;
 234 }
 235
 236 static void __init zswap_entry_cache_destroy(void)
 237 {
 238         kmem_cache_destroy(zswap_entry_cache);
 239 }
 240
 241 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
 242 {
 243         struct zswap_entry *entry;
 244         entry = kmem_cache_alloc(zswap_entry_cache, gfp);
 245         if (!entry)
 246                 return NULL;
 247         entry->refcount = 1;
 248         RB_CLEAR_NODE(&entry->rbnode);
 249         return entry;
 250 }
 251
 252 static void zswap_entry_cache_free(struct zswap_entry *entry)
 253 {
 254         kmem_cache_free(zswap_entry_cache, entry);
 255 }
 256
 257 /*********************************
 258 * rbtree functions
 259 **********************************/
 260 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
 261 {
 262         struct rb_node *node = root->rb_node;
 263         struct zswap_entry *entry;
 264
 265         while (node) {
 266                 entry = rb_entry(node, struct zswap_entry, rbnode);
 267                 if (entry->offset > offset)
 268                         node = node->rb_left;
 269                 else if (entry->offset < offset)
 270                         node = node->rb_right;
 271                 else
 272                         return entry;
 273         }
 274         return NULL;
 275 }
 276
 277 /*
 278  * In the case that a entry with the same offset is found, a pointer to
 279  * the existing entry is stored in dupentry and the function returns -EEXIST
 280  */
 281 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
 282                         struct zswap_entry **dupentry)
 283 {
 284         struct rb_node **link = &root->rb_node, *parent = NULL;
 285         struct zswap_entry *myentry;
 286
 287         while (*link) {
 288                 parent = *link;
 289                 myentry = rb_entry(parent, struct zswap_entry, rbnode);
 290                 if (myentry->offset > entry->offset)
 291                         link = &(*link)->rb_left;
 292                 else if (myentry->offset < entry->offset)
 293                         link = &(*link)->rb_right;
 294                 else {
 295                         *dupentry = myentry;
 296                         return -EEXIST;
 297                 }
 298         }
 299         rb_link_node(&entry->rbnode, parent, link);
 300         rb_insert_color(&entry->rbnode, root);
 301         return 0;
 302 }
 303
 304 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
 305 {
 306         if (!RB_EMPTY_NODE(&entry->rbnode)) {
 307                 rb_erase(&entry->rbnode, root);
 308                 RB_CLEAR_NODE(&entry->rbnode);
 309         }
 310 }
 311
 312 /*
 313  * Carries out the common pattern of freeing and entry's zpool allocation,
 314  * freeing the entry itself, and decrementing the number of stored pages.
 315  */
 316 static void zswap_free_entry(struct zswap_entry *entry)
 317 {
 318         zpool_free(entry->pool->zpool, entry->handle);
 319         zswap_pool_put(entry->pool);
 320         zswap_entry_cache_free(entry);
 321         atomic_dec(&zswap_stored_pages);
 322         zswap_update_total_size();
 323 }
 324
 325 /* caller must hold the tree lock */
 326 static void zswap_entry_get(struct zswap_entry *entry)
 327 {
 328         entry->refcount++;
 329 }
 330
 331 /* caller must hold the tree lock
 332 * remove from the tree and free it, if nobody reference the entry
 333 */
 334 static void zswap_entry_put(struct zswap_tree *tree,
 335                         struct zswap_entry *entry)
 336 {
 337         int refcount = --entry->refcount;
 338
 339         BUG_ON(refcount < 0);
 340         if (refcount == 0) {
 341                 zswap_rb_erase(&tree->rbroot, entry);
 342                 zswap_free_entry(entry);
 343         }
 344 }
 345
 346 /* caller must hold the tree lock */
 347 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 348                                 pgoff_t offset)
 349 {
 350         struct zswap_entry *entry;
 351
 352         entry = zswap_rb_search(root, offset);
 353         if (entry)
 354                 zswap_entry_get(entry);
 355
 356         return entry;
 357 }
 358
 359 /*********************************
 360 * per-cpu code
 361 **********************************/
 362 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
 363
 364 static int zswap_dstmem_prepare(unsigned int cpu)
 365 {
 366         u8 *dst;
 367
 368         dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
 369         if (!dst) {
 370                 pr_err("can't allocate compressor buffer\n");
 371                 return -ENOMEM;
 372         }
 373         per_cpu(zswap_dstmem, cpu) = dst;
 374         return 0;
 375 }
 376
 377 static int zswap_dstmem_dead(unsigned int cpu)
 378 {
 379         u8 *dst;
 380
 381         dst = per_cpu(zswap_dstmem, cpu);
 382         kfree(dst);
 383         per_cpu(zswap_dstmem, cpu) = NULL;
 384
 385         return 0;
 386 }
 387
 388 static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
 389 {
 390         struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
 391         struct crypto_comp *tfm;
 392
 393         if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
 394                 return 0;
 395
 396         tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
 397         if (IS_ERR_OR_NULL(tfm)) {
 398                 pr_err("could not alloc crypto comp %s : %ld\n",
 399                        pool->tfm_name, PTR_ERR(tfm));
 400                 return -ENOMEM;
 401         }
 402         *per_cpu_ptr(pool->tfm, cpu) = tfm;
 403         return 0;
 404 }
 405
 406 static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
 407 {
 408         struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
 409         struct crypto_comp *tfm;
 410
 411         tfm = *per_cpu_ptr(pool->tfm, cpu);
 412         if (!IS_ERR_OR_NULL(tfm))
 413                 crypto_free_comp(tfm);
 414         *per_cpu_ptr(pool->tfm, cpu) = NULL;
 415         return 0;
 416 }
 417
 418 /*********************************
 419 * pool functions
 420 **********************************/
 421
 422 static struct zswap_pool *__zswap_pool_current(void)
 423 {
 424         struct zswap_pool *pool;
 425
 426         pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
 427         WARN_ON(!pool);
 428
 429         return pool;
 430 }
 431
 432 static struct zswap_pool *zswap_pool_current(void)
 433 {
 434         assert_spin_locked(&zswap_pools_lock);
 435
 436         return __zswap_pool_current();
 437 }
 438
 439 static struct zswap_pool *zswap_pool_current_get(void)
 440 {
 441         struct zswap_pool *pool;
 442
 443         rcu_read_lock();
 444
 445         pool = __zswap_pool_current();
 446         if (!pool || !zswap_pool_get(pool))
 447                 pool = NULL;
 448
 449         rcu_read_unlock();
 450
 451         return pool;
 452 }
 453
 454 static struct zswap_pool *zswap_pool_last_get(void)
 455 {
 456         struct zswap_pool *pool, *last = NULL;
 457
 458         rcu_read_lock();
 459
 460         list_for_each_entry_rcu(pool, &zswap_pools, list)
 461                 last = pool;
 462         if (!WARN_ON(!last) && !zswap_pool_get(last))
 463                 last = NULL;
 464
 465         rcu_read_unlock();
 466
 467         return last;
 468 }
 469
 470 /* type and compressor must be null-terminated */
 471 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
 472 {
 473         struct zswap_pool *pool;
 474
 475         assert_spin_locked(&zswap_pools_lock);
 476
 477         list_for_each_entry_rcu(pool, &zswap_pools, list) {
 478                 if (strcmp(pool->tfm_name, compressor))
 479                         continue;
 480                 if (strcmp(zpool_get_type(pool->zpool), type))
 481                         continue;
 482                 /* if we can't get it, it's about to be destroyed */
 483                 if (!zswap_pool_get(pool))
 484                         continue;
 485                 return pool;
 486         }
 487
 488         return NULL;
 489 }
 490
 491 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
 492 {
 493         struct zswap_pool *pool;
 494         char name[38]; /* 'zswap' + 32 char (max) num + \0 */
 495         gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
 496         int ret;
 497
 498         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 499         if (!pool) {
 500                 pr_err("pool alloc failed\n");
 501                 return NULL;
 502         }
 503
 504         /* unique name for each pool specifically required by zsmalloc */
 505         snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
 506
 507         pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
 508         if (!pool->zpool) {
 509                 pr_err("%s zpool not available\n", type);
 510                 goto error;
 511         }
 512         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
 513
 514         strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
 515         pool->tfm = alloc_percpu(struct crypto_comp *);
 516         if (!pool->tfm) {
 517                 pr_err("percpu alloc failed\n");
 518                 goto error;
 519         }
 520
 521         ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
 522                                        &pool->node);
 523         if (ret)
 524                 goto error;
 525         pr_debug("using %s compressor\n", pool->tfm_name);
 526
 527         /* being the current pool takes 1 ref; this func expects the
 528          * caller to always add the new pool as the current pool
 529          */
 530         kref_init(&pool->kref);
 531         INIT_LIST_HEAD(&pool->list);
 532
 533         zswap_pool_debug("created", pool);
 534
 535         return pool;
 536
 537 error:
 538         free_percpu(pool->tfm);
 539         if (pool->zpool)
 540                 zpool_destroy_pool(pool->zpool);
 541         kfree(pool);
 542         return NULL;
 543 }
 544
 545 static __init struct zswap_pool *__zswap_pool_create_fallback(void)
 546 {
 547         if (!crypto_has_comp(zswap_compressor, 0, 0)) {
 548                 if (!strcmp(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT)) {
 549                         pr_err("default compressor %s not available\n",
 550                                zswap_compressor);
 551                         return NULL;
 552                 }
 553                 pr_err("compressor %s not available, using default %s\n",
 554                        zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
 555                 param_free_charp(&zswap_compressor);
 556                 zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
 557         }
 558         if (!zpool_has_pool(zswap_zpool_type)) {
 559                 if (!strcmp(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT)) {
 560                         pr_err("default zpool %s not available\n",
 561                                zswap_zpool_type);
 562                         return NULL;
 563                 }
 564                 pr_err("zpool %s not available, using default %s\n",
 565                        zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
 566                 param_free_charp(&zswap_zpool_type);
 567                 zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
 568         }
 569
 570         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
 571 }
 572
 573 static void zswap_pool_destroy(struct zswap_pool *pool)
 574 {
 575         zswap_pool_debug("destroying", pool);
 576
 577         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
 578         free_percpu(pool->tfm);
 579         zpool_destroy_pool(pool->zpool);
 580         kfree(pool);
 581 }
 582
 583 static int __must_check zswap_pool_get(struct zswap_pool *pool)
 584 {
 585         return kref_get_unless_zero(&pool->kref);
 586 }
 587
 588 static void __zswap_pool_release(struct work_struct *work)
 589 {
 590         struct zswap_pool *pool = container_of(work, typeof(*pool), work);
 591
 592         synchronize_rcu();
 593
 594         /* nobody should have been able to get a kref... */
 595         WARN_ON(kref_get_unless_zero(&pool->kref));
 596
 597         /* pool is now off zswap_pools list and has no references. */
 598         zswap_pool_destroy(pool);
 599 }
 600
 601 static void __zswap_pool_empty(struct kref *kref)
 602 {
 603         struct zswap_pool *pool;
 604
 605         pool = container_of(kref, typeof(*pool), kref);
 606
 607         spin_lock(&zswap_pools_lock);
 608
 609         WARN_ON(pool == zswap_pool_current());
 610
 611         list_del_rcu(&pool->list);
 612
 613         INIT_WORK(&pool->work, __zswap_pool_release);
 614         schedule_work(&pool->work);
 615
 616         spin_unlock(&zswap_pools_lock);
 617 }
 618
 619 static void zswap_pool_put(struct zswap_pool *pool)
 620 {
 621         kref_put(&pool->kref, __zswap_pool_empty);
 622 }
 623
 624 /*********************************
 625 * param callbacks
 626 **********************************/
 627
 628 /* val must be a null-terminated string */
 629 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
 630                              char *type, char *compressor)
 631 {
 632         struct zswap_pool *pool, *put_pool = NULL;
 633         char *s = strstrip((char *)val);
 634         int ret;
 635
 636         if (zswap_init_failed) {
 637                 pr_err("can't set param, initialization failed\n");
 638                 return -ENODEV;
 639         }
 640
 641         /* no change required */
 642         if (!strcmp(s, *(char **)kp->arg))
 643                 return 0;
 644
 645         /* if this is load-time (pre-init) param setting,
 646          * don't create a pool; that's done during init.
 647          */
 648         if (!zswap_init_started)
 649                 return param_set_charp(s, kp);
 650
 651         if (!type) {
 652                 if (!zpool_has_pool(s)) {
 653                         pr_err("zpool %s not available\n", s);
 654                         return -ENOENT;
 655                 }
 656                 type = s;
 657         } else if (!compressor) {
 658                 if (!crypto_has_comp(s, 0, 0)) {
 659                         pr_err("compressor %s not available\n", s);
 660                         return -ENOENT;
 661                 }
 662                 compressor = s;
 663         } else {
 664                 WARN_ON(1);
 665                 return -EINVAL;
 666         }
 667
 668         spin_lock(&zswap_pools_lock);
 669
 670         pool = zswap_pool_find_get(type, compressor);
 671         if (pool) {
 672                 zswap_pool_debug("using existing", pool);
 673                 list_del_rcu(&pool->list);
 674         } else {
 675                 spin_unlock(&zswap_pools_lock);
 676                 pool = zswap_pool_create(type, compressor);
 677                 spin_lock(&zswap_pools_lock);
 678         }
 679
 680         if (pool)
 681                 ret = param_set_charp(s, kp);
 682         else
 683                 ret = -EINVAL;
 684
 685         if (!ret) {
 686                 put_pool = zswap_pool_current();
 687                 list_add_rcu(&pool->list, &zswap_pools);
 688         } else if (pool) {
 689                 /* add the possibly pre-existing pool to the end of the pools
 690                  * list; if it's new (and empty) then it'll be removed and
 691                  * destroyed by the put after we drop the lock
 692                  */
 693                 list_add_tail_rcu(&pool->list, &zswap_pools);
 694                 put_pool = pool;
 695         }
 696
 697         spin_unlock(&zswap_pools_lock);
 698
 699         /* drop the ref from either the old current pool,
 700          * or the new pool we failed to add
 701          */
 702         if (put_pool)
 703                 zswap_pool_put(put_pool);
 704
 705         return ret;
 706 }
 707
 708 static int zswap_compressor_param_set(const char *val,
 709                                       const struct kernel_param *kp)
 710 {
 711         return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
 712 }
 713
 714 static int zswap_zpool_param_set(const char *val,
 715                                  const struct kernel_param *kp)
 716 {
 717         return __zswap_param_set(val, kp, NULL, zswap_compressor);
 718 }
 719
 720 static int zswap_enabled_param_set(const char *val,
 721                                    const struct kernel_param *kp)
 722 {
 723         if (zswap_init_failed) {
 724                 pr_err("can't enable, initialization failed\n");
 725                 return -ENODEV;
 726         }
 727
 728         return param_set_bool(val, kp);
 729 }
 730
 731 /*********************************
 732 * writeback code
 733 **********************************/
 734 /* return enum for zswap_get_swap_cache_page */
 735 enum zswap_get_swap_ret {
 736         ZSWAP_SWAPCACHE_NEW,
 737         ZSWAP_SWAPCACHE_EXIST,
 738         ZSWAP_SWAPCACHE_FAIL,
 739 };
 740
 741 /*
 742  * zswap_get_swap_cache_page
 743  *
 744  * This is an adaption of read_swap_cache_async()
 745  *
 746  * This function tries to find a page with the given swap entry
 747  * in the swapper_space address space (the swap cache).  If the page
 748  * is found, it is returned in retpage.  Otherwise, a page is allocated,
 749  * added to the swap cache, and returned in retpage.
 750  *
 751  * If success, the swap cache page is returned in retpage
 752  * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
 753  * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
 754  *     the new page is added to swapcache and locked
 755  * Returns ZSWAP_SWAPCACHE_FAIL on error
 756  */
 757 static int zswap_get_swap_cache_page(swp_entry_t entry,
 758                                 struct page **retpage)
 759 {
 760         bool page_was_allocated;
 761
 762         *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
 763                         NULL, 0, &page_was_allocated);
 764         if (page_was_allocated)
 765                 return ZSWAP_SWAPCACHE_NEW;
 766         if (!*retpage)
 767                 return ZSWAP_SWAPCACHE_FAIL;
 768         return ZSWAP_SWAPCACHE_EXIST;
 769 }
 770
 771 /*
 772  * Attempts to free an entry by adding a page to the swap cache,
 773  * decompressing the entry data into the page, and issuing a
 774  * bio write to write the page back to the swap device.
 775  *
 776  * This can be thought of as a "resumed writeback" of the page
 777  * to the swap device.  We are basically resuming the same swap
 778  * writeback path that was intercepted with the frontswap_store()
 779  * in the first place.  After the page has been decompressed into
 780  * the swap cache, the compressed version stored by zswap can be
 781  * freed.
 782  */
 783 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
 784 {
 785         struct zswap_header *zhdr;
 786         swp_entry_t swpentry;
 787         struct zswap_tree *tree;
 788         pgoff_t offset;
 789         struct zswap_entry *entry;
 790         struct page *page;
 791         struct crypto_comp *tfm;
 792         u8 *src, *dst;
 793         unsigned int dlen;
 794         int ret;
 795         struct writeback_control wbc = {
 796                 .sync_mode = WB_SYNC_NONE,
 797         };
 798
 799         /* extract swpentry from data */
 800         zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
 801         swpentry = zhdr->swpentry; /* here */
 802         zpool_unmap_handle(pool, handle);
 803         tree = zswap_trees[swp_type(swpentry)];
 804         offset = swp_offset(swpentry);
 805
 806         /* find and ref zswap entry */
 807         spin_lock(&tree->lock);
 808         entry = zswap_entry_find_get(&tree->rbroot, offset);
 809         if (!entry) {
 810                 /* entry was invalidated */
 811                 spin_unlock(&tree->lock);
 812                 return 0;
 813         }
 814         spin_unlock(&tree->lock);
 815         BUG_ON(offset != entry->offset);
 816
 817         /* try to allocate swap cache page */
 818         switch (zswap_get_swap_cache_page(swpentry, &page)) {
 819         case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
 820                 ret = -ENOMEM;
 821                 goto fail;
 822
 823         case ZSWAP_SWAPCACHE_EXIST:
 824                 /* page is already in the swap cache, ignore for now */
 825                 put_page(page);
 826                 ret = -EEXIST;
 827                 goto fail;
 828
 829         case ZSWAP_SWAPCACHE_NEW: /* page is locked */
 830                 /* decompress */
 831                 dlen = PAGE_SIZE;
 832                 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
 833                                 ZPOOL_MM_RO) + sizeof(struct zswap_header);
 834                 dst = kmap_atomic(page);
 835                 tfm = *get_cpu_ptr(entry->pool->tfm);
 836                 ret = crypto_comp_decompress(tfm, src, entry->length,
 837                                              dst, &dlen);
 838                 put_cpu_ptr(entry->pool->tfm);
 839                 kunmap_atomic(dst);
 840                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
 841                 BUG_ON(ret);
 842                 BUG_ON(dlen != PAGE_SIZE);
 843
 844                 /* page is up to date */
 845                 SetPageUptodate(page);
 846         }
 847
 848         /* move it to the tail of the inactive list after end_writeback */
 849         SetPageReclaim(page);
 850
 851         /* start writeback */
 852         __swap_writepage(page, &wbc, end_swap_bio_write);
 853         put_page(page);
 854         zswap_written_back_pages++;
 855
 856         spin_lock(&tree->lock);
 857         /* drop local reference */
 858         zswap_entry_put(tree, entry);
 859
 860         /*
 861         * There are two possible situations for entry here:
 862         * (1) refcount is 1(normal case),  entry is valid and on the tree
 863         * (2) refcount is 0, entry is freed and not on the tree
 864         *     because invalidate happened during writeback
 865         *  search the tree and free the entry if find entry
 866         */
 867         if (entry == zswap_rb_search(&tree->rbroot, offset))
 868                 zswap_entry_put(tree, entry);
 869         spin_unlock(&tree->lock);
 870
 871         goto end;
 872
 873         /*
 874         * if we get here due to ZSWAP_SWAPCACHE_EXIST
 875         * a load may happening concurrently
 876         * it is safe and okay to not free the entry
 877         * if we free the entry in the following put
 878         * it it either okay to return !0
 879         */
 880 fail:
 881         spin_lock(&tree->lock);
 882         zswap_entry_put(tree, entry);
 883         spin_unlock(&tree->lock);
 884
 885 end:
 886         return ret;
 887 }
 888
 889 static int zswap_shrink(void)
 890 {
 891         struct zswap_pool *pool;
 892         int ret;
 893
 894         pool = zswap_pool_last_get();
 895         if (!pool)
 896                 return -ENOENT;
 897
 898         ret = zpool_shrink(pool->zpool, 1, NULL);
 899
 900         zswap_pool_put(pool);
 901
 902         return ret;
 903 }
 904
 905 /*********************************
 906 * frontswap hooks
 907 **********************************/
 908 /* attempts to compress and store an single page */
 909 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
 910                                 struct page *page)
 911 {
 912         struct zswap_tree *tree = zswap_trees[type];
 913         struct zswap_entry *entry, *dupentry;
 914         struct crypto_comp *tfm;
 915         int ret;
 916         unsigned int dlen = PAGE_SIZE, len;
 917         unsigned long handle;
 918         char *buf;
 919         u8 *src, *dst;
 920         struct zswap_header *zhdr;
 921
 922         if (!zswap_enabled || !tree) {
 923                 ret = -ENODEV;
 924                 goto reject;
 925         }
 926
 927         /* reclaim space if needed */
 928         if (zswap_is_full()) {
 929                 zswap_pool_limit_hit++;
 930                 if (zswap_shrink()) {
 931                         zswap_reject_reclaim_fail++;
 932                         ret = -ENOMEM;
 933                         goto reject;
 934                 }
 935         }
 936
 937         /* allocate entry */
 938         entry = zswap_entry_cache_alloc(GFP_KERNEL);
 939         if (!entry) {
 940                 zswap_reject_kmemcache_fail++;
 941                 ret = -ENOMEM;
 942                 goto reject;
 943         }
 944
 945         /* if entry is successfully added, it keeps the reference */
 946         entry->pool = zswap_pool_current_get();
 947         if (!entry->pool) {
 948                 ret = -EINVAL;
 949                 goto freepage;
 950         }
 951
 952         /* compress */
 953         dst = get_cpu_var(zswap_dstmem);
 954         tfm = *get_cpu_ptr(entry->pool->tfm);
 955         src = kmap_atomic(page);
 956         ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
 957         kunmap_atomic(src);
 958         put_cpu_ptr(entry->pool->tfm);
 959         if (ret) {
 960                 ret = -EINVAL;
 961                 goto put_dstmem;
 962         }
 963
 964         /* store */
 965         len = dlen + sizeof(struct zswap_header);
 966         ret = zpool_malloc(entry->pool->zpool, len,
 967                            __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM,
 968                            &handle);
 969         if (ret == -ENOSPC) {
 970                 zswap_reject_compress_poor++;
 971                 goto put_dstmem;
 972         }
 973         if (ret) {
 974                 zswap_reject_alloc_fail++;
 975                 goto put_dstmem;
 976         }
 977         zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
 978         zhdr->swpentry = swp_entry(type, offset);
 979         buf = (u8 *)(zhdr + 1);
 980         memcpy(buf, dst, dlen);
 981         zpool_unmap_handle(entry->pool->zpool, handle);
 982         put_cpu_var(zswap_dstmem);
 983
 984         /* populate entry */
 985         entry->offset = offset;
 986         entry->handle = handle;
 987         entry->length = dlen;
 988
 989         /* map */
 990         spin_lock(&tree->lock);
 991         do {
 992                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
 993                 if (ret == -EEXIST) {
 994                         zswap_duplicate_entry++;
 995                         /* remove from rbtree */
 996                         zswap_rb_erase(&tree->rbroot, dupentry);
 997                         zswap_entry_put(tree, dupentry);
 998                 }
 999         } while (ret == -EEXIST);
1000         spin_unlock(&tree->lock);
1001
1002         /* update stats */
1003         atomic_inc(&zswap_stored_pages);
1004         zswap_update_total_size();
1005
1006         return 0;
1007
1008 put_dstmem:
1009         put_cpu_var(zswap_dstmem);
1010         zswap_pool_put(entry->pool);
1011 freepage:
1012         zswap_entry_cache_free(entry);
1013 reject:
1014         return ret;
1015 }
1016
1017 /*
1018  * returns 0 if the page was successfully decompressed
1019  * return -1 on entry not found or error
1020 */
1021 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1022                                 struct page *page)
1023 {
1024         struct zswap_tree *tree = zswap_trees[type];
1025         struct zswap_entry *entry;
1026         struct crypto_comp *tfm;
1027         u8 *src, *dst;
1028         unsigned int dlen;
1029         int ret;
1030
1031         /* find */
1032         spin_lock(&tree->lock);
1033         entry = zswap_entry_find_get(&tree->rbroot, offset);
1034         if (!entry) {
1035                 /* entry was written back */
1036                 spin_unlock(&tree->lock);
1037                 return -1;
1038         }
1039         spin_unlock(&tree->lock);
1040
1041         /* decompress */
1042         dlen = PAGE_SIZE;
1043         src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
1044                         ZPOOL_MM_RO) + sizeof(struct zswap_header);
1045         dst = kmap_atomic(page);
1046         tfm = *get_cpu_ptr(entry->pool->tfm);
1047         ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
1048         put_cpu_ptr(entry->pool->tfm);
1049         kunmap_atomic(dst);
1050         zpool_unmap_handle(entry->pool->zpool, entry->handle);
1051         BUG_ON(ret);
1052
1053         spin_lock(&tree->lock);
1054         zswap_entry_put(tree, entry);
1055         spin_unlock(&tree->lock);
1056
1057         return 0;
1058 }
1059
1060 /* frees an entry in zswap */
1061 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1062 {
1063         struct zswap_tree *tree = zswap_trees[type];
1064         struct zswap_entry *entry;
1065
1066         /* find */
1067         spin_lock(&tree->lock);
1068         entry = zswap_rb_search(&tree->rbroot, offset);
1069         if (!entry) {
1070                 /* entry was written back */
1071                 spin_unlock(&tree->lock);
1072                 return;
1073         }
1074
1075         /* remove from rbtree */
1076         zswap_rb_erase(&tree->rbroot, entry);
1077
1078         /* drop the initial reference from entry creation */
1079         zswap_entry_put(tree, entry);
1080
1081         spin_unlock(&tree->lock);
1082 }
1083
1084 /* frees all zswap entries for the given swap type */
1085 static void zswap_frontswap_invalidate_area(unsigned type)
1086 {
1087         struct zswap_tree *tree = zswap_trees[type];
1088         struct zswap_entry *entry, *n;
1089
1090         if (!tree)
1091                 return;
1092
1093         /* walk the tree and free everything */
1094         spin_lock(&tree->lock);
1095         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1096                 zswap_free_entry(entry);
1097         tree->rbroot = RB_ROOT;
1098         spin_unlock(&tree->lock);
1099         kfree(tree);
1100         zswap_trees[type] = NULL;
1101 }
1102
1103 static void zswap_frontswap_init(unsigned type)
1104 {
1105         struct zswap_tree *tree;
1106
1107         tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
1108         if (!tree) {
1109                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1110                 return;
1111         }
1112
1113         tree->rbroot = RB_ROOT;
1114         spin_lock_init(&tree->lock);
1115         zswap_trees[type] = tree;
1116 }
1117
1118 static struct frontswap_ops zswap_frontswap_ops = {
1119         .store = zswap_frontswap_store,
1120         .load = zswap_frontswap_load,
1121         .invalidate_page = zswap_frontswap_invalidate_page,
1122         .invalidate_area = zswap_frontswap_invalidate_area,
1123         .init = zswap_frontswap_init
1124 };
1125
1126 /*********************************
1127 * debugfs functions
1128 **********************************/
1129 #ifdef CONFIG_DEBUG_FS
1130 #include <linux/debugfs.h>
1131
1132 static struct dentry *zswap_debugfs_root;
1133
1134 static int __init zswap_debugfs_init(void)
1135 {
1136         if (!debugfs_initialized())
1137                 return -ENODEV;
1138
1139         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1140         if (!zswap_debugfs_root)
1141                 return -ENOMEM;
1142
1143         debugfs_create_u64("pool_limit_hit", S_IRUGO,
1144                         zswap_debugfs_root, &zswap_pool_limit_hit);
1145         debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1146                         zswap_debugfs_root, &zswap_reject_reclaim_fail);
1147         debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1148                         zswap_debugfs_root, &zswap_reject_alloc_fail);
1149         debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1150                         zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1151         debugfs_create_u64("reject_compress_poor", S_IRUGO,
1152                         zswap_debugfs_root, &zswap_reject_compress_poor);
1153         debugfs_create_u64("written_back_pages", S_IRUGO,
1154                         zswap_debugfs_root, &zswap_written_back_pages);
1155         debugfs_create_u64("duplicate_entry", S_IRUGO,
1156                         zswap_debugfs_root, &zswap_duplicate_entry);
1157         debugfs_create_u64("pool_total_size", S_IRUGO,
1158                         zswap_debugfs_root, &zswap_pool_total_size);
1159         debugfs_create_atomic_t("stored_pages", S_IRUGO,
1160                         zswap_debugfs_root, &zswap_stored_pages);
1161
1162         return 0;
1163 }
1164
1165 static void __exit zswap_debugfs_exit(void)
1166 {
1167         debugfs_remove_recursive(zswap_debugfs_root);
1168 }
1169 #else
1170 static int __init zswap_debugfs_init(void)
1171 {
1172         return 0;
1173 }
1174
1175 static void __exit zswap_debugfs_exit(void) { }
1176 #endif
1177
1178 /*********************************
1179 * module init and exit
1180 **********************************/
1181 static int __init init_zswap(void)
1182 {
1183         struct zswap_pool *pool;
1184         int ret;
1185
1186         zswap_init_started = true;
1187
1188         if (zswap_entry_cache_create()) {
1189                 pr_err("entry cache creation failed\n");
1190                 goto cache_fail;
1191         }
1192
1193         ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
1194                                 zswap_dstmem_prepare, zswap_dstmem_dead);
1195         if (ret) {
1196                 pr_err("dstmem alloc failed\n");
1197                 goto dstmem_fail;
1198         }
1199
1200         ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1201                                       "mm/zswap_pool:prepare",
1202                                       zswap_cpu_comp_prepare,
1203                                       zswap_cpu_comp_dead);
1204         if (ret)
1205                 goto hp_fail;
1206
1207         pool = __zswap_pool_create_fallback();
1208         if (!pool) {
1209                 pr_err("pool creation failed\n");
1210                 goto pool_fail;
1211         }
1212         pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1213                 zpool_get_type(pool->zpool));
1214
1215         list_add(&pool->list, &zswap_pools);
1216
1217         frontswap_register_ops(&zswap_frontswap_ops);
1218         if (zswap_debugfs_init())
1219                 pr_warn("debugfs initialization failed\n");
1220         return 0;
1221
1222 pool_fail:
1223         cpuhp_remove_state_nocalls(CPUHP_MM_ZSWP_POOL_PREPARE);
1224 hp_fail:
1225         cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
1226 dstmem_fail:
1227         zswap_entry_cache_destroy();
1228 cache_fail:
1229         /* if built-in, we aren't unloaded on failure; don't allow use */
1230         zswap_init_failed = true;
1231         zswap_enabled = false;
1232         return -ENOMEM;
1233 }
1234 /* must be late so crypto has time to come up */
1235 late_initcall(init_zswap);
1236
1237 MODULE_LICENSE("GPL");
1238 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1239 MODULE_DESCRIPTION("Compressed cache for swap pages");